Ultrasonic image recognition method and system based on deep convolutional neural network
Through deep convolutional neural networks and Faster R-CNN models, the problem of ultrasound image interpretation relying on operator experience is solved, and efficient recognition of ultrasound images and automatic classification of standard sections are achieved, which improves diagnostic efficiency and accuracy. It is particularly suitable for cardiac and obstetrics and gynecology ultrasound examinations.
Patent Information
- Application Number
- CN202510966577.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-14
- Publication Date
- 2025-10-17
AI Technical Summary
The interpretation of ultrasound images is highly dependent on the operator's experience, resulting in significant differences between different operators. Manually obtaining standard sections is time-consuming and error-prone, making standardization and automation difficult to achieve, affecting diagnostic efficiency and quality.
Using deep convolutional neural networks and Faster R-CNN models, through image preprocessing and training, efficient recognition and section classification of ultrasound images are achieved. This includes normalization, image enhancement, grayscale conversion, histogram equalization, and Gaussian difference function processing. It also combines feature extraction, region proposal, and classification and regression networks to reduce human error.
It improves the accuracy and efficiency of ultrasound examinations, especially in cardiac and obstetrics and gynecology ultrasound examinations, realizes the automatic recognition and diagnosis of standard sections, reduces human errors, and improves the accuracy and consistency of diagnosis.
Smart Images
Figure CN120808029A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of image recognition, and more particularly to an ultrasonic image recognition method and system based on a deep convolutional neural network. BACKGROUND
[0002] Currently, the interpretation of ultrasonic images highly depends on the experience and skills of operators, especially when obtaining standard sections. There can be significant differences between different operators. Standard sections are key images in ultrasonic examinations for accurately evaluating specific anatomical structures, such as the left ventricular long-axis section in cardiac ultrasound or the fetal head circumference section in obstetric and gynecologic ultrasound. The accurate acquisition of these sections is crucial for disease diagnosis and treatment. In traditional methods, operators need to manually adjust the probe to obtain standard sections, which not only consumes time but also easily leads to inaccurate section acquisition due to insufficient experience or improper operation, thereby affecting the diagnostic results. In addition, manual operation is difficult to realize standardization and automation, which limits the efficiency and quality of ultrasonic examinations.
[0003] Therefore, it is of great significance to design an ultrasonic image recognition method and system based on a deep convolutional neural network for improving the accuracy and efficiency of ultrasonic examinations. SUMMARY
[0004] Therefore, the present application provides an ultrasonic image recognition method and system based on a deep convolutional neural network, which realizes efficient recognition and section classification of ultrasonic images through a deep convolutional neural network and a Faster R-CNN model.
[0005] To achieve the above purpose, the present application adopts the following technical solution: an ultrasonic image recognition method based on a deep convolutional neural network, comprising: acquiring an ultrasonic image and pre-processing the ultrasonic image; constructing a Faster R-CNN target detection network model and training it; inputting the pre-processed ultrasonic image into the trained target detection network model to output a model recognition result; outputting the section type of the ultrasonic image according to the model recognition result.
[0006] Preferably, the pre-processing of the image includes: the ultrasonic image is sequentially subjected to normalization processing, image enhancement, grayscale processing, histogram equalization processing, and Gaussian difference function processing; the grayscale-processed image, the equalized image, and the Gaussian difference function-processed image are fused respectively to obtain a corresponding three-channel image; the three-channel image is used as a pre-processed image to constitute a pre-processing data set.
[0007] Preferably, the preprocessed data set is divided into a training set, a validation set and a test set.
[0008] Preferably, the Faster R-CNN target detection network model comprises a feature extraction network, a region proposal network, a region of interest pooling and a classification and regression network. The preprocessed data set is input into the Faster R-CNN target detection network model, The feature extraction network extracts feature maps in the preprocessed data set; The region proposal network generates anchor points in the feature maps, predicts the offset of the bounding box according to the anchor points containing the target, adjusts the position and size of the anchor points, filters out high-confidence candidate regions, and removes regions with large overlap through non-maximum suppression to retain high-quality candidate regions; The region of interest pooling maps the generated high-quality candidate regions back to the feature maps, and extracts features for each candidate region; The classification and regression network classifies and regresses the bounding box for the candidate region; The classification and regression network performs non-maximum suppression on the classification and regression results, removes bounding boxes with large overlap, retains the final detection results, and outputs the detected target class and bounding box position.
[0009] Preferably, the training set is used to train the target detection network model; a loss function and an optimization algorithm are selected during the training process; training parameters are set; the weights of the model are continuously adjusted through the back propagation algorithm, so that the target detection network model can fit the training data.
[0010] Preferably, during the training process, the validation set is used to validate the model and evaluate the performance indicators of the model; according to the validation results, the model is optimized and adjusted to prevent overfitting.
[0011] Preferably, after training is completed, the test set is used to test the model and evaluate the final performance of the model.
[0012] Preferably, an ultrasonic image recognition system based on a deep convolutional neural network comprises: An image processing module for acquiring an ultrasonic image and pre-processing the ultrasonic image; A model construction module for constructing and training a Faster R-CNN target detection network model; An image recognition module for inputting the pre-processed ultrasonic image into the trained target detection network model and outputting a model recognition result; An output module for outputting the cross-sectional type of the ultrasonic image according to the model recognition result.
[0013] Compared with the prior art, the ultrasonic image recognition method and system based on a deep convolutional neural network are provided, the efficient recognition and section classification of ultrasonic images are realized through the deep convolutional neural network and the Faster R-CNN model, the pre-processing and training steps ensure the accurate learning of the model on the image features, the section type is output in real time, the diagnostic efficiency and accuracy are improved, the human error is reduced, and the method is especially suitable for complex ultrasonic examination scenes such as the heart and obstetrics, and has important clinical application value. BRIEF DESCRIPTION OF DRAWINGS
[0014] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of the provided drawings.
[0015] Figure 1 The method flowchart provided by the present application is shown in the figure. DETAILED DESCRIPTION
[0016] The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.
[0017] As shown in the figure, the present application discloses an ultrasonic image recognition method based on a deep convolutional neural network, comprising: Figure 1 acquiring an ultrasonic image and pre-processing the ultrasonic image; constructing a Faster R-CNN target detection network model and training the model; inputting the pre-processed ultrasonic image into the trained target detection network model, and outputting the model recognition result; outputting the section type of the ultrasonic image according to the model recognition result. Specifically, the pre-processing process of the image comprises:
[0018] the ultrasonic image is sequentially subjected to normalization processing, image enhancement, grayscale processing, histogram equalization processing and Gaussian difference function processing; the grayscale processed image, the equalized image and the Gaussian difference function processed image are fused respectively to obtain the corresponding three-channel image; the grayscale processed image, the equalized image and the Gaussian difference function processed image are fused respectively to obtain the corresponding three-channel image; The three-channel image constitutes a pre-processing data set as a pre-processing image.
[0019] Specifically, the pre-processing data set is divided into a training set, a validation set and a test set.
[0020] Specifically, the Faster R-CNN target detection network model comprises a feature extraction network, a region proposal network, a region of interest pooling and a classification and regression network. The pre-processing data set is input into the Faster R-CNN target detection network model, The feature extraction network extracts feature maps in the pre-processing data set. The region proposal network generates anchor points in the feature maps, predicts the offset of the bounding box according to the anchor points containing the target, adjusts the position and size of the anchor points, filters out high-confidence candidate regions, and removes regions with large overlaps through non-maximum suppression to retain high-quality candidate regions. The region of interest pooling maps the generated high-quality candidate regions back to the feature maps, and extracts features for each candidate region. The classification and regression network is used for classification and bounding box regression of the candidate regions. The classification and regression network performs non-maximum suppression on the classification and regression results, removes bounding boxes with large overlaps, retains the final detection results, and outputs the detected target class and bounding box position.
[0021] Specifically, the training set is used to train the target detection network model; a loss function and an optimization algorithm are selected during the training process; training parameters are set; the weights of the model are continuously adjusted through the back propagation algorithm, so that the target detection network model can fit the training data.
[0022] Specifically, during the training process, the validation set is used to validate the model and evaluate the performance indicators of the model; according to the validation result, the model is optimized and adjusted to prevent overfitting.
[0023] Specifically, after training, the test set is used to test the model and evaluate the final performance of the model.
[0024] In a specific embodiment of the present application, the pre-processed image frame is input into the trained target detection network model. The model automatically extracts the features of the image and judges whether the image is a standard section.
[0025] If the model identifies the current frame as a standard section, the section type (such as left ventricular long axis section, etc.) is output; if the current frame is not a standard section, a position deviation vector is further calculated.
[0026] If the class outputted by the model is a standard section type (e.g. left ventricular long axis section), the current frame is determined as a standard section.
[0027] If the class outputted by the model is background or other non-standard section type, the current frame is determined as a non-standard section.
[0028] The candidate regions can be further filtered according to a confidence threshold (e.g. 0.9), only the candidate regions with confidence higher than the threshold are determined as standard sections.
[0029] Further, for non-standard sections, the model can calculate a position deviation vector by comparing the feature differences between the current image and the standard section. The position deviation vector can be represented as a two-dimensional vector (Δx, Δy), where Δx represents the deviation in the horizontal direction and Δy represents the deviation in the vertical direction. The calculation method can be based on the matching of feature points or the calculation of image similarity. For example, by detecting the position difference between the key feature points (e.g. the position of the mitral valve in echocardiography) in the current image and the corresponding feature points in the standard section, the position deviation vector is calculated.
[0030] Specifically, an ultrasound image recognition system based on a deep convolutional neural network comprises: An image processing module for acquiring an ultrasound image and pre-processing the ultrasound image; A model construction module for constructing and training a Faster R-CNN target detection network model; An image recognition module for inputting the pre-processed ultrasound image into the trained target detection network model and outputting a model recognition result; An output module for outputting the section type of the ultrasound image according to the model recognition result.
[0031] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other.
[0032] The above description of the disclosed embodiments enables a person skilled in the art to implement or use the present application. Various modifications to the embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. An ultrasound image recognition method based on deep convolutional neural network, characterized in that: include: Acquiring an ultrasound image and preprocessing the ultrasound image; Build and train the Faster R-CNN target detection network model; Input the preprocessed ultrasound image into the trained target detection network model and output the model recognition result; Determine whether the image is a standard section based on the model recognition results.
2. The ultrasound image recognition method based on deep convolutional neural network according to claim 1, characterized in that: The process of image preprocessing includes: performing normalization processing, image enhancement, grayscale conversion, histogram equalization processing, and Gaussian difference function processing on the ultrasound image in sequence; The grayscale image, the equalized image, and the image processed by the Gaussian difference function are fused to obtain the corresponding three-channel image; The three-channel images serve as preprocessed images to form a preprocessed data set.
3. The ultrasound image recognition method based on deep convolutional neural network according to claim 2, characterized in that: The preprocessed dataset is divided into a training set, a validation set, and a test set.
4. The ultrasound image recognition method based on deep convolutional neural network according to claim 1, characterized in that: The Faster R-CNN target detection network model includes: feature extraction network, region proposal network, region of interest pooling, and classification and regression network; Input the preprocessed dataset into the Faster R-CNN target detection network model, Extract feature maps from the preprocessed dataset through a feature extraction network; The region proposal network generates anchors in the feature map, predicts the offset of the bounding box based on the anchor containing the target, adjusts the position and size of the anchor, filters out high-confidence candidate regions, and removes large overlapping regions through non-maximum suppression to retain high-quality candidate regions; Region of interest pooling maps the generated high-quality candidate regions back to the feature map and performs feature extraction on each candidate region; Use classification and regression networks to classify candidate regions and perform bounding box regression; The classification and regression network performs non-maximum suppression on the classification and regression results, removes bounding boxes with large overlaps, retains the final detection results, and outputs the detected target category and bounding box position.
5. The ultrasound image recognition method based on deep convolutional neural network according to claim 3, characterized in that: Use the training set to train the target detection network model; select the loss function and optimization algorithm during the training process; set the training parameters; and continuously adjust the model weights through the backpropagation algorithm so that the target detection network model can fit the training data.
6. The ultrasound image recognition method based on deep convolutional neural network according to claim 5, characterized in that: During the training process, the validation set is used to validate the model and evaluate the model's performance indicators; based on the validation results, the model is optimized and adjusted to prevent overfitting.
7. The ultrasound image recognition method based on deep convolutional neural network according to claim 6, characterized in that: After training is completed, the model is tested using the test set to evaluate the final performance of the model.
8. An ultrasound image recognition system based on a deep convolutional neural network, characterized in that: include: Image processing module: used to acquire ultrasound images and pre-process them; Model building module: used to build and train the Faster R-CNN target detection network model; Image recognition module: used to input the preprocessed ultrasound image into the trained target detection network model and output the model recognition results; Output module: used to output the section type of the ultrasound image based on the model recognition results.